Discriminative training of multi-state barge-in models for speech processing

1. The method of training a barge-in-model for speech processing, the method comprising:

receiving an input having at least one speech segment and at least one non-speech segment;

establishing a restriction of recognizing only speech states during speech segments of the input and non-speech states during non-speech segments of the input;

generating a hypothesis lattice by allowing any sequence of speech Hidden Markov Models (HMMs) and non-speech HMMs;

generating a reference lattice by only allowing speech HMMs for at least one speech segment and non-speech HMMs for at least one non-speech segment, wherein different iterations of training generates at least one different reference lattice and at least one reference transcription; and

employing the generated reference lattice as the barge-in-model for speech processing.

2. The method of claim 1 , wherein the barge-in-model uses a maximum likelihood (ML) training steps and a maximum mutual information (MMI) training steps.

3. The method of claim 1 , wherein the at least one reference lattice that is redefined at each iteration of training.

4. The method of claim 1 , wherein speech is recognized as one word and non-speech is recognized as being an additional word.

5. The method of claim 1 , further comprising recognizing non-speech states during speech segments but only with a high penalty to minimize such occurrences.

6. The method of claim 1 , wherein the restriction is slightly relaxed to allow for non-speech HMM alignments due to mislabeling at a beginning or ending of the speech segments.